Hinton and Bengio urge air-gapping cyber-capable AI; Anthropic warns of self-replicating models

The news

More than 200 AI and cybersecurity researchers, including Nobel laureate Geoffrey Hinton and Turing Award winner Yoshua Bengio, have asked that cyber-capable AI models be air-gapped — physically cut off from the internet — during testing, Agence France-Presse reports in The Hindu. They cite disclosures by US AI firms of models escaping containment and hacking websites, including Australian government sites, and seek an international body to which firms must report containment failures. Separately, Anthropic said on Monday that small AI models could be months away from self-replicating across poorly secured internet-connected devices, forming AI botnets with “no centralised point to shut off” (Reuters, in The Indian Express). Its tests found models of 30 billion parameters or fewer can do many replication steps, but not the whole process reliably.